Medical Diagnostic Tool with Neural Model Trained Through Machine Learning for Predicting Coronary Disease from ECG Signals

PRIORITY

CLAIM

The present application is a continuation-in-part of U.S. patent application Ser. No. 18/160,613, filed Jan. 27, 2023, which is a continuation of U.S. patent application Ser. No. 17/382,909, filed Jul. 22, 2021, now U.S. Pat. No. 11,568,991, issued Jan. 31, 2023, which are incorporated herein by reference in their entirety, and which claims priority to U.S. Provisional Patent Application No. 63/055,603, titled MEDICAL DIAGNOSTIC TOOL WITH NEURAL MODEL TRAINED THROUGH MACHINE LEARNING FOR PREDICTING CORONARY DISEASE FROM ECG SIGNALS, filed Jul. 23, 2020.

BACKGROUND

Approximately seven million Americans undergo emergency cardiac triage each year and the length of time to reach a final coronary diseases diagnosis is 8 to 72 hours. Also, approximately 40% of patients receive unnecessary coronary catherization, and 55-85% patients are considered to be over treated. Despite these measures, 2-6% of sick patients are discharged without treatment. Nearly 4 million cardiac catheterizations are performed annually in United States hospitals alone, with one in ten discharges having undergone coronary arteriography. Of these procedures, it is found that over 60% of them find patients with non-obstructive coronary artery disease-which is surprising as this is an invasive test reserved for the highest risk patients. Tests to indicate coronary disease, and thus need for cardiac catheterization, are generally accurate to an average of about 85%; meaning patients are sent for invasive catheterization with significant uncertainty of the need for such procedure. The innovation presented herein increases this accuracy and therefore provides for better estimate of procedure need, and it does so within seconds, and for a fraction of the cost of the hours-long testing currently used.

SUMMARY

The present invention is directed, in various aspects, to a medical diagnostic tool that help doctors (or other applicable health officials as the case may be) rapidly identify clinically significant cardiovascular disease and sort patients that need further testing from those that can be safely discharged. The software-based diagnostic tool can be trained through machine learning to enable users to better analyze and diagnose a patient's ECG signal. The software for the tool can be used with already cleared/approved ECG devices. The tool's software is able to accept input from any ECG device, regardless of device gain and sample frequency. The tool employs non-invasive technology that uses, for example, standard 12-lead ECG data to predict the presence, location, and severity of atherosclerotic cardiovascular disease (ACVD) and coronary artery disease (CAD). The results can be provided within second, as opposed to hours for current cardiac triage techniques. Plus, preliminary testing shows that the results are 95% accurate and the accuracy rate will improve with more data because the tool is trained through machine learning. In one general aspect, the present invention is directed to a diagnostic tool that includes: a sensor for capturing at least one biosignal produced by a patient's heart; and a computer device that implements a deep neural network that is trained iteratively through machine learning to generate a prediction about a heart condition of the patient. After the deep neural network is trained, the computer device is configured to: convert the at least one biosignal to a multi-dimensional input matrix for the deep neural network generated from a number (N) of biosignals captured by the sensor, where each of the N biosignals is at least “T” seconds, and where T is at least one second and N is greater than 1; and process the multi-dimensional input matrix through the deep neural network. The output of the deep neural network from processing the multi-dimensional input matrix corresponds to the prediction about the heart condition of the patient. In another general aspect, the present invention is directed to a method including: training, with a computer system, iteratively through machine learning, a deep neural network to make a prediction about a heart condition of a patient; and after training the deep neural network: capturing, with a sensor, at least one biosignal produced by a patient's heart; converting, by the computer system, the at least one biosignal to an input matrix for the deep neural network, wherein the input matrix includes a multi-dimensional matrix generated from a number (N) of biosignals of at least “T” seconds (where Tis at least one second and N is greater than 1); and processing, by the computer system, the multi-dimensional input matrix through the deep neural network. The output of the deep neural network from processing the multi-dimensional input matrix through the deep neural network corresponds to the prediction about the heart condition of the patient. In another general aspect, the present invention is directed to a computer system including one or more processor cores, wherein the one or more processor cores are configured to: train, iteratively through machine learning, a deep neural network to make a prediction about a heart condition of a patient; and after training the deep neural network: receive, from a sensor, at least one ECG signal produced by a patient's heart and captured by the sensor; convert the at least one ECG signal to an input matrix for the deep neural network, wherein the input matrix includes a multi-dimensional (or multi-axis) matrix generated from N ECG signals of at least T seconds, where T is at least one second and N is greater than 1; and process the input matrix through the deep neural network, wherein an output of the deep neural network from processing the input matrix through the deep neural network corresponds to the prediction about the heart condition of the patient. The present invention, in various embodiments, provide many benefits over prior art systems. First, as described herein, the diagnostic tool can accurately and rapidly detect coronary diseases of a patient based on, for example, ECG signals of the patient. Second, the diagnostic tool can provide those accurate and rapid results even if not all twelve ECG leads are used on the patient. These and other benefits of the present invention will be apparent from the description to follow. FIGURES Various aspects of the present invention are described herein by way of example in connection with the following figures. FIG. 1 is diagram of a medical diagnostic tool according to various aspects of the present invention. FIGS. 2 and 3 depict a process for training and validating a neural model of the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIG. 4 depicts outputs of the neural model of the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIGS. 5 A-B and 6 A-B depict visualizations generated by the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIG. 7 is a diagram that depicts the neural model of the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIG. 8 depicts a densely connected convolutional network that can be used as part of the neural model of the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIG. 9 is a block diagram of the computer device of the medical diagnostic tool of FIG. 1 according to various aspects of the present invention. FIG. 10 A is a diagram of a block of a residual network (ResNet) according to various embodiments of the present invention. FIG. 10 B is a diagram of a ResNet according to various embodiments of the present invention. FIG. 11 B is a diagram of a vision transformer (ViT) according to various embodiments of the present invention. FIG. 11 B is a diagram of a transformer encoder of the ViT of FIG. 11 A according to various embodiments of the present invention. DESCRIPTION The present invention is directed, in various aspects, to a medical diagnostic tool 100 for identifying clinically significant cardiovascular disease in a patient 101 . In various aspects, the diagnostic tool 100 includes, as shown in FIG. 1 , at least one sensor configured to capture biosignals associated with the patient. For example, the sensor 102 of FIG. 1 includes one or more leads of an ECG machine. Although the non-limiting aspect of FIG. 1 includes an ECG machine with one or more leads, it shall be appreciated that the sensor 102 can include any number of devices configured to capture any number of biosignals associated with the patient. For example, in other non-limiting aspects, the sensor 102 can include an electroencephalogram (EEG) machine. In still other non-limiting aspects, the sensor 102 can include a wearable device (e.g., a smart watch, a smart ring, a smart phone, or smart glasses, etc.) configured to generate either ECG or EEG signals. HEARTio. Furthermore, the sensor 102 can include any number of leads. According to the non-limiting aspect of FIG. 1 , the sensor 102 could be a conventional 12-lead ECG machine, although other types of machines with any number of leads could be used. The sensor 102 can be communicably coupled to a back-end computer device 104 , sometimes referred to herein as a “HEARTio computer.” The computer device 104 includes modules 108 trained through machine learning to identify the existence (or non-existence) of one or more clinically significant cardiovascular diseases based on the ECG data from a patient. The HEARTio computer 104 stores and executes a software program 108 that analyses a patient's biosignals (e.g., 12-lead ECG data) and outputs a predicted level, severity, and localization of coronary stenosis or other coronary diseases that the software is trained, through machine learning, to identify. The ECG data can be input, for example, to computer device 104 in the following ways: (1) Direct input of digital data points in a patient dataset via the hospital infrastructure; (2) Manual input of a digital signal file through an email client; (3) Manual input of a digital signal data via a dashboard provided by the computer device 104 ; and/or (4) Manual input of an biosignal waveform (non-digital signal such as a pdf or image) via a dashboard. For this last option, the computer device 104 can transform the non-digital signal into a set of a digital signal data. The computer device 104 could have a direct wired network (e.g., an Ethernet) or wireless (e.g., WIFI) connection to the computer device 104 . Also, the computer device 104 could be on the cloud, in which case the biosignal data from a patient is uploaded to a database on the cloud, which database the computer device 104 can access to analyze the patient's biosignal data to make the diagnosis. The software program 108 can include one or more neural models and, after input of biosignal data, the data is pushed through the neural model(s), sometimes referred to herein as a “HEARTio neural model” as shown in FIG. 1 , implemented by the software 108 of the HEARTio computer 104 . In that connection, the HEARTio neural model 108 can be a fixed graph that has a very specific set of instructions for how to process the input data. Then HEARTio computer 104 can separate the biosignal data into more than, for example, 100 distinct layers of the graph (or neural network) by applying a specific set of mathematical operations to the data exiting the preceding layer in the neural graph. The final layer of the HEARTio neural model 108 preferably provides an array of probabilities for the question being asked (Is this biosignal representative of coronary stenosis? How severe is this coronary stenosis? What is the location of the coronary stenosis? etc.) The difference between a normal software program and the HEARTio neural model 108 is that the set of instructions for the HEARTio neural model 108 is not determined by a physician, programmer, or expert, but by the HEARTio computer 104 itself. A typical software program implements a known flow of information through a fixed set of code to perform processes on the data, but those processes are designed by the programmer. With the diagnostic tool 100 of the present invention, the determination of these instructions is done algorithmically, meaning the HEARTio computer 104 itself determines what mathematical operations are used to operate on data exiting one layer of the graph and being delivered to the next layer. A training set of data is used to train the neural model 108 using, for example, supervised, unsupervised, and/or semi-supervised training. The graph is initially set up to use a randomly generated set of instructions to deliver the next layer of data from the preceding layer, with the first layer being the “training set” of biosignal data, singularly, for each patient in the “training set”. The biosignal data in the training set is pushed through the various layers of the graph, with its mathematical operations between each layer, and then, after the last layer, the initial set of answers are recorded. In supervised training, the answers are known, as part of the iterative training, the answers produced by the model 108 can be compared to the actual answers/results associated with the training data. If the experimental results and the expected results are far away in magnitude, then the instructions representing the mathematical operations between the layers is adjusted (done by the HEARTio computer 104 according to an algorithm) to move the model 108 closer to a set of operations that may offer better results at the end of the last layer. If the distance between the real answer and the answer computed by the combination of the all the operations between the numerous layers in the graph is minimal, possibly only the operations between lower levels are adjusted. This process is repeated, over and over, until the congruence between the experimental results and the expected results are satisfactory, at which point the model of mathematical operations between the layers can then be applied on additional patient biosignal data. After training has concluded, the final graph (set of mathematical operations acting between the various layers of the model) is fixed and ready to be used to process biosignal data that was not included in the training set. The process of inputting new data to the neural model 108 is called “testing”. The main advantages to allowing the computer 104 to determine the optimal set of instructions are that a) it can revise the instructions and evaluate the results much more quickly than with human involvement, b) it can use a much more complicated instruction set than can be envisioned by humans, and c) it can update the instructions with the acquisition of new data to get, theoretically, more and more accurate results with more and more experience. Thus, the biosignal neural model 108 will adapt to novel data, leading to more accurate analysis of biosignal data with respect to coronary stenosis or other monitored—for conditions as the case may be. FIG. 2 depicts the training and validation processes 200 in more detail. The patient cohort is split into two groups: a “training set” 202 and a “testing set” 204 . Generally accepted deep learning practice is to utilize greater than 90% of the patient cohort for the training set 202 and the remainder to validate the accuracy statistics of the model 206 . The programmer may also manually set the hyperparameters, for the training, which are parameters unrelated to the training set but will impact the training environment (e.g., learning rate, label smoothing, dropout rate). Once the patient cohort is split and before training begins, the model randomizes the model 206 parameters it will refine to build the statistical dependencies between input and output. Over time the model 206 improves upon these parameters utilizing a process known as gradient descent, in which each of the parameters searches for local minimums of a loss function (the measure of how different the prediction is from the actual value). Training the model 206 is iterative (as shown in FIG. 3 ). The number of epochs (a measure of how many forward and backward passes the model 206 makes as part of the iterative gradient descent training), and the number of steps per epoch (a measure of how many data points are evaluated within the epoch) may be set by the programmer. In each epoch, a training set is randomly pulled from the training cohort 204 . The model 206 then makes predictions about the patient based on patient biosignals, trying to identify important parameters that make the correct diagnosis. Once the forward pass (forward propagation) is complete, the predictions are sent to the loss function, which compares the predictions against the actual patient conditions and determines the overall error. This error then gets back-propagated through the model 206 and helps adjusts parameters to correct for the model 206 error. After all the epochs are complete, the final model 206 is considered “trained” and that model is passed to testing. This trained model 206 is then given the previously defined “testing set”, which it has never seen, to predict patient conditions. These predictions are compared to the actual patient outcomes and these results are reported as model 206 performance (usually in terms of sensitivity, Positive Predictive Value (PPV), and F-1 score). The longer the duration of the input 208 fed into the HEARTio neural model 206 generally the better. Preferably, the duration of the input 208 should be at least one second. In various aspects, the input 208 ( FIG. 4 ) that is fed into HEARTio neural model 206 is more than 5 seconds and less than 15 seconds, such as ten seconds, of standard ECG signals (or other biosignals as the case may be) and can be anywhere from one to twelve leads. The model 206 output 210 ( FIG. 4 ) can include, for example, in six different but related areas: beat classification, rhythm classification, Myocardial Infarction (MI) classification, Major Adverse Cardiovascular Events (MACE) prediction, and CAD severity classification. Of course, the output 210 can include any other information associated with the heart of the patient 101 ( FIG. 1 ) as determined by the model 206 . Unlike most current biosignal interpretation software, which simply monitors basic beats, rhythm, and MI detection, the HEARTio computer 104 ( FIG. 1 ), with its HEARTio neural model 206 , can be trained to detect, for example, 17 different and highly clinically relevant beat types, 17 different and highly clinically relevant rhythm types, and five different and highly clinically relevant MI types. The HEARTio computer 104 ( FIG. 1 ) can provide this improvement through in MACE prediction, and CAD severity classification prediction. The HEARTio neural model 206 can be trained to predict a patient's risk for a MACE, determine that patient's current level of CAD, and localize stenosis in the patient's coronary arteries. The outputs that the HEARTio neural model 206 can be trained, in various aspects, to detect are listed in Table 1 below. TABLE 1 Beats Rhythm Possible Labels Possible Labels Aberrated atrial premature beat Atrial fibrillation Atrial escape beat Atrial flutter Atrial premature beat First heart block Fusion of paced and normal beat Idioventricular rhythm Fusion of ventricular and normal beat Nodal (A-V junctional) rhythm Left bundle branch block beat Noise Nodal (junctional) escape beat Normal sinus rhythm Nodal (junctional) premature beat Paced rhythm Non-conducted P-wave (blocked APC) Pre-excitation (WPW) Normal beat Second heart block Paced beat Supraventricular tachyarrhythmia Premature ventricular contraction Ventricular bigeminy Right bundle branch block beat Ventricular flutter Supraventricular premature or Ventricular tachycardia ectopic beat (atrial or nodal) Unclassifiable beat Ventricular trigeminy Ventricular escape beat Atrial bigeminy Ventricular flutter wave Sinus bradycardia MI MACE Risk Severity Localization Possible Possible Possible Vessel Blockage Labels Labels Labels Percentage Acute MI None Low/Mild LAD % Prior MI MI Moderate LCX % NSTEMI Stroke Severe LM % Prior CABG Syncope RCA % Prior PCI The HEARTio neural model 206 could also be trained to detect or score Coronary artery calcium (CAC); Absolute Agatston score (continuous); Multivessel FFR (Continuous); Multivessel Binary Fractional Flow Reserve (FFR) (Binary); and/or instantaneous wave-free ratio (IFR). The HEARTio computer device 104 ( FIG. 1 ) may also be programmed to generate graphical displays that display the results of its analysis, which can be accessed via a display 108 ( FIG. 1 ) communicably coupled to the computer device 104 ( FIG. 1 ). For example, the graphical output displays could be displayed on a monitor of the HEARTio computer device or on another device, e.g., the user device 108 shown in FIG. 1 , that is in data communication with the HEARTio computer device 104 ( FIG. 1 ). For example, the HEARTio computer device 104 ( FIG. 1 ) could include a web server that serves the results via web pages to the user device 108 ( FIG. 1 ). FIGS. 5 A-B and 6 A-B show example graphical output displays 500 , 510 that can be generated by the HEARTio computer device 104 ( FIG. 1 ). FIG. 5 A is an example graphical display 500 that could be displayed, for example, to an emergency medicine physician or a triage nurse. The example 500 of FIG. 5 A can include a preliminary assessment 502 of the patient's heart, an ischemia assessment 504 include an estimated risk of CAD, and a MACE assessment 506 , all of which are determined by the trained model 206 ( FIGS. 2 - 4 ). The display 500 of FIG. 5 A can be used by an ED physician. It includes fields that convey: whether the patient's beats and rhythms are normal or not and whether MI is present 502 ; an ischemia assessment 504 , such as whether there is CAD risk or whether vessel diseases was detected; and a MACE assessment 506 , including the risk of 12 month MI, 12 month stroke, and 12 month syncope. The biosignal (e.g., ECG, EEG, etc.) waveforms 508 at the bottom of FIG. 5 A show most important biosignal leads that contribute to the severity distinction as determined by the HEARTio neural map. A second graphical display, such as the display 510 of FIG. 5 B , can include fields that show biosignal morphology 512 , important biosignal leads 514 , whether abnormal beats were detected 516 , whether abnormal rhythms were detected 518 , occlusion amounts (e.g., percentages) of various arteries 520 , as well as biosignals from various leads (e.g., the important leads) 522 . For example, as shown in FIG. 5 B , the output visualization can show the percent lesion blockage 520 within each of the 4 major coronary arteries. If the blockages are low, moderate, or high, they can be shaded differently (e.g., green, yellow, red). FIGS. 6 A and 6 B shows other graphical visualizations 600 , 610 that can be provided by the computer system 104 ( FIG. 1 ). The visualization 600 of FIG. 6 A shows the predicted levels of 4 different blood tests 602 , 604 , 606 , 608 that the HEARTio model 206 ( FIGS. 2 - 4 ) has generated outputs for. Once again, the graphical visualization 600 can use color coding to visually communicate information to a using technician. For example, if the levels are low the bars can be green; if they are moderate it will be yellow; and if they are high it will be red (or some other color coding). The visualization 610 of FIG. 6 B shows the level of ejection fraction. The higher the level the heart graphic fills up. If the EF is low, the color in the graphic can be yellow and, if even lower, red (or some other color coding). In various aspects, the input to HEARTio neural model 206 ( FIGS. 2 - 4 ) is a tensor with 4 axes: batch, signal, lead, and channel. The batch axis refers to the total number of biosignal that are being processed at the same time; if a single biosignal is being evaluated then the batch axis size would be 1. The signal axis represents the discrete numerical signal value for 10 seconds of time; in various aspects, the size of this axis is nonvariable and is 1000. The lead axis represents the standard 12 leads (e.g., the conventional I, II, II, aVL, aVF, aVL, V1, V2, V3, V4, V5, V6 leads) that include a resting biosignal; and this axis can be nonvariable in certain aspects. The channel axis is preferably nonvariable as well and set to 1. This axis is used as a convenience for convolutional processing. Overall a biosignal can be represented by 12,000 double (float32) values. Many biosignal algorithms are based on the concept of rule-based analysis, meaning that are rules set in stone determined by expert opinion or rudimentary analysis. This type of analysis usually requires the manual input of demographic information such as sex, age, race, family history. The HEARTio neural model does not require any further input than the biosignal represented in tensor form and can be blind to any other information. The tensor form biosignal, for example, can include a conventional ECG with anywhere from 1 to 12 leads organized as shown in Table 2 below. Alternatively, the tensor form biosignal can include a conventional EEG with anywhere from 1 to 19 leads. If a lead does not exist in the original signal it is replaced by a series of zeros. If the tensor form biosignal was extracted from an image or PDF there might a series of zeros in different parts of various leads. TABLE 2 ECG leads and their respective index in the lead axis Lead Lead Axis Index I 0 II 1 III 2 aVR 3 aVL 4 aVF 5 V1 6 V2 7 V3 8 V4 9 V5 10 V6 11 X 0 Y 1 Z 2 To create a biosignal in tensor form the following steps may be followed. First, clip a biosignal such that it represents (for example) 10 seconds and is in the form N×M, where Mis the number of the leads and N is the number of samples. N is equal to the sampling rate of the biosignal multiplied by 10. Second, using fast Fourier transform (FFT), resample N to 1000, in effect reducing the sampling rate to 100 Hz. Third, bandpass Butterworth filter with a passband starting at 2 Hz and extending to 40 Hz. This will try to make sure only the information within this band of frequencies are retained while others are removed. Fourth, each filtered signal can be scaled by the following formula: f ⁡ ( x ) = 2 ⁢ x - min ⁡ ( x ) max ⁡ ( x ) - min ⁡ ( x ) - 1 This will make every value between −1 and +1. Fifth, any NaN value can be converted to 0 . And sixth, each signal can be detrended such that the isoelectrical portions of the biosignal are 0. After this level of preprocessing there still might be signals, or leads of signals, that do not have a reasonable amount or degree of usable information. In that case, there is a series of steps that are used in training the neural model that can also be used to disregard the signals without usable information. The outputs of HEARTio neural model can fit into the following major archetypes, where each index of the output array represents a class or category: 1. Reconstruction: Array is the form of a reconstructed biosignal. 2. Scaled-Continuous: Values in this output array are continuous but scaled to be between 0 and 1. 3. Multi-Class: Each output array can only contain 1 value of 1, all other values are 0. There is only 1 class that be present. 4. Multi-Label: Each output array can have multiple values of 1. Each class can be present or absent. Table 3 below shows an example macro listing of output of the HEARTio neural model and the shaped associated with each output. TABLE 3 Output Name Length Type Reconstruction ECG {1000, 12, 1} Reconstruction Coronary Blockage % 4 Scaled-Continuous MI & History 5 Multi-Label Abnormal Beats 17 Multi-Label Abnormal Rhythms 17 Multi-Label SPECT 20 Scaled-Continuous MACE Risk 4 Multi-Class CAD Severity-Overall 3 Multi-Class CAD Severity-Leads 3 Multi-Class Coronary Binary Blockage 4 Multi-Label Blood Test Predictions 4 Scaled-Continuous Complications 4 Multi-Label Ejection Fraction 1 Scaled-Continuous Disease 1 Multi-Label Conduction Disorders 11 Multi-Label Hypertrophy 5 Multi-Label Location of MI 14 Multi-Label Morphology 15 Multi-Label ST-T Changes 13 Multi-Label Ejection Fraction Binary 1 Multi-Label FFR 1 Scaled-Continuous FFR Binary 1 Multi-Label Table 4, in Appendix A hereto, shows the indexes within each major category according to various aspects of the present invention. In various aspects, the basis of HEARTio neural model 206 ( FIG. 2 ) is the development and training of three distinct models that work in concert to provide information and thoroughly analyze biosignals. The steps of the HEARTio neural model 206 ( FIG. 2 ), once trained, are as follows, with reference to FIG. 7 : 1. A biosignal 701 is input and processed through an encoder 704 which performs a lossy compression 2. The minimized representation of this lossy compression, called the latent space 708 , serves as input to two layers: a concatenation layer 716 (used later on), and a decoder 706 3. The decoder 706 converts the latent space representation 708 back into the tensorform biosignal 710 4. The tensorform biosignal 710 is input into a Dense network 712 to produce a feature vector 714 5. The feature vector 714 and the latent space 708 are concatenated to form one feature vector 716 , which is used as input to a classifier 718 6. Outputs 720 from the classifier 718 are collected at the end The encoder 704 and decoder 706 form an autoencoder 702 . The autoencoder 702 can use information about the biosignal and any extra leads to remove non-significant noise from the biosignal in its reconstruction. The autoencoder 702 can also detect with biosignal leads are empty and, in various aspects, fill them in accordingly using information from existing leads. For example, where there is just one existing lead, the autoencoder 702 can generate the 11 other, missing leads so that there are 12 signals (e.g., leads I, II, II, aVL, aVF, aVL, V1, V2, V3, V4, V5, V6) to be used by the dense network 712 (or ResNet or Vit as the case may be). The autoencoder can also be trained to generate any Frank lead (e.g., X, Y, Z leads), which measure cardiac activity, which allows for vectorcardiography analysis. The autoencoder 702 can also standardize the biosignals such that there is no artifact that is only related to one type of biosignal lead or patient. In various aspects, the encoder 704 can include approximately 13.5 million total parameters, with approximately 99.97% of them being trainable parameters (with the remainder being non-trainable parameters), as shown in Table 5 at Appendix B. Table 6 at Appendix C shows, in one aspect, a detailed description of each type, output shape, and number of parameters associated with each later of the decoder of FIG. 7 . Table 7 below shows, in one aspect, a detailed description of each type, output shape, and number of parameters associated with each layer of the overall autoencoder 702 . The encoder 704 can be configured to perform a lossy compression of the at least one biosignal captured by the sensor 102 ( FIG. 1 ), thereby producing a latent space representation 708 as an output of the lossy compression. The decoder 706 receives the latent space representation 708 as an input from the encoder 704 and can be configured to convert the latent space representation 708 into a tensorform signal output 710 provided to the DenseNet neural network 712 . In other embodiments, other types of deep neural networks could be used, such as in place of the DenseNet neural network 712 , such as a residual neural network (ResNet) or a vision transformer (ViT), as explained further below. TABLE 7 Model: “Autoencoder” Layer (type) Output Shape Param # input 3 (InputLayer) (None, 1000, 12, 1) 0 Encoder (Model) (None, 1024) 13451936 Decoder (Model) (None, 1000, 12, 1) 13324801 lead adjuster 1 (LeadAdjusted) (None, 1000, 12, 1) 1 Total params: 26,776,738 Trainable params: 26,746,018 Non-trainable params: 30,720 The autoencoder may also be trained to predict the next X seconds of Y leads given X seconds of N number of leads. For example, if 5 seconds from one or more ECG leads are captured from the patient, the autoencoder 702 , once trained, can generate the next X seconds for 1 to 15 leads (e.g., the 12 conventional leads plus the 3 Frank leads). The DenseNet 712 shown in FIG. 7 may be implemented with a form of a deep neural network 712 , including an input layer 802 , an output layer 808 , and one or more hidden layers 804 , 806 between the input layer 802 and output layers 808 (see FIG. 8 ), where each layer includes at least node, and directed arcs between nodes. An output value for each node can be computed by the computer system according to an activation function for the node, where the inputs to the activation function for a node are weighted outputs of the node(s) having directed arcs to the node for which the activation value is being computed. The weights for the arcs and any applicable bias values can be learned parameters that are learned through the training. In various aspects, the dense network 712 shown in FIG. 7 can be implemented with a densely connected neural network (DenseNet) 712 . In such a DenseNet 712 , each layer obtains additional inputs from all preceding layers and passes own its own feature-maps to all subsequent layers. Concatenation 716 during forward propagation can be used, such as shown in FIG. 8 , wherein the circle c's denote concatenation. Each layer 802 , 804 , 806 , 808 can receive a “collective knowledge” from all preceding layers of the neural network 712 . For example, after receiving the tensorform signal output 710 from the decoder 706 of the autoencoder 702 , the DenseNet 712 can be configured generate a feature vector 714 based, at least in part, on the tensorform signal 710 . The DenseNet 712 of FIG. 8 can include an input layer 802 , a concatenation layer 804 , a classifier layer 806 , and an output layer 808 . According to the non-limiting aspect of FIG. 8 , the concatenation layer 804 can be configured to receive the feature vector from the input layer 802 and the latent space representation into a concatenated feature vector However, it shall be appreciated that in other non-limiting aspects, the DenseNet 712 can include these and any other layers configured to generate outputs that can produce meaningful predictions regarding the heart condition of the patient. In that case, the network 712 can be relatively thinner and compact, resulting in high computational efficiency and memory efficiency. Table 8 in Appendix D provides a detailed description of each type, output shape, number of parameters, and connections associated with each layer of the dense network according to one aspect of the present invention. Table 9 in Appendix E provides a detailed description of the full neural network, including the encoder, decoder and dense network, according to various aspects of the present invention. Other types of deep neural networks could be used in place of, or in addition to, the DenseNet 712 according to various embodiments. For example, a ResNet and/or a ViT could be employed. A ResNet primarily consists of convolutional layers that perform feature extraction, in this case, for example, with respect to the ECG signal 710 , particularly in the form of a biosignal image, as described herein. The convoltional layers detect low-level features and gradually build up to detect higher-level features. The ResNet can comprise one or more residual connections 10 , as shown in FIGS. 10 A and 10 B . FIG. 10 A shows an example block of a ResNet with a residual connection 10 that skips a layer 11 of the network and enables the gradient to flow more easily during training, allowing for the construction of very deep networks. Instead of learning the target mapping, the network can learn a residual, which is added to the input to obtain the output. FIG. 10 B is a diagram of a ResNet with the block of FIG. 10 A according to various embodiments. As shown in the example of FIG. 10 B , the ResNet can comprise one or more convolutional layers 13 , as well as one or more pooling layers 15 interspersed to reduce the spatial dimensions and control overfitting. As mentioned previously, a ViT could also be used in various embodiments. FIG. 11 A illustrates an example of a ViT artchitecture and FIG. 11 B illustrates an example of the transformer encoder thereof, according to various embodiments. With reference to FIGS. 11 A and 11 B , a ViT is a transformer that breaks down an input image, in this case, for example, the biosignal image described herein, into a series of patches 21 , serializes each patch into a vector, and maps it to a smaller dimension with a single matrix multiplication. These vector embeddings 18 are then processed by a transformer encoder 20 , such as shown FIG. 11 B . The transformer encoder 20 can comprise alternating layers of multihead self-attentuation 22 and multilayer perceptrons (MLP) 26 . A layernorm 26 can be applied before every block and with residual connections 28 after every block. The MPL can contain two layers with a GELU non-linearity. The final layer, MLP head 23 , can provide an output of the transformer. An application of softmax on this output can provide classification labels. Like the DenseNet 712 shown in FIG. 7 , the output of the ResNet or ViT, as the case may be, can comprise a feature vector 714 that is input to a fully connected layer 716 to map the feature vector 714 to the output classes 720 . As mentioned above, the HEARTio neural model (e.g., DenseNet, ResNet, or ViT) can be trained to work with all twelve leads of a standard 12-lead ECG machine or with less than all of the leads downs to just one lead. During training, the following adjust adjustment process can be used to adjust the training data to train the HEARTio neural model to work with 1 to 12 leads. If the training data has 12 leads: X1% chance of leaving the signal as is X2% chance of randomly choosing 1 lead X3% chance of randomly choosing up to 11 leads X4% chance of removing 75% of each lead where X1+X2+X3+X4=100% If input ECG training data has greater than 1 but less than 12 leads Y1% chance of leaving the signal as is Y2% chance of randomly choosing 1 lead Y3% chance of randomly choosing up to N−1 leads where Y1+Y2+Y3=100%; and If input ECG has only 1 lead, leave signal as is. The autoencoder 702 ( FIG. 7 ) may employ a reconstruction loss function to make that the autoencoder's 702 ( FIG. 7 ) reconstruction remains as faithful to the input biosignal. The reconstruction loss function may be based on a squared difference between the fast Fourier transform (FFT) between the input lead 701 ( FIG. 7 ) (if available) and the output lead 710 ( FIG. 7 ) from the autoencoder 702 ( FIG. 7 ). Values with 0s in the biosignal can be ignored and not considered. The loss can be average per input signal. The loss function set forth below can be used by any output that is of the scaled-continuous variety. Input biosignals that contain −1 can be masked and ignored. Each of the gold-truth outputs can be stratified into, for example, 25 bins that are evenly distributed between 0 and 1. Each bin can be cycled through and the mean squared log error (MSLE) between the predicted and true arrays that correspond to a particular bin can be calculated. This can be performed for every index found in the scaled-continuous output. The loss can be averaged among all indexes and bins. def weighed_stats_MSLE (y_true, y_pred): mask_true=K. cast (K. not_equal (y_true,−1), K. float x ( )) num_true=K.sum (mask_true) nbins=25 indices=t f. histogram_fixed_width_bins (mask_true*y_true, ¥ [0.0, 1.0], nbins=nbins) num_class=y_true. get shape ( ). as list ( ) [−1] unique, =tf.unique (K. flatten (indices)) unique=K.sum (K. cast (K. ones like (unique), K. float x ( ))) first_log=K.log (K.clip (mask_true*y_pred, K. epsilon ( ) None)+1.) second_log=K.log (K.clip (mask_true*y_true, K.epsilon ( ) None)+1.) mse=K. square (first_log−second_log) loss=0.0 for j in range (nbins): inds=mask_true*K. cast (K. equal (indices, j), K. float x ( )) for i in range (num_class): loss+=1.0/(unique*num_class)*K.sum (¥ inds [:, i] *mse [:, i])/(K.sum (inds [:, i])+K. epsilon ( )) return loss Also, the loss function set forth below can be used by the any output that is of the multi-class variety. Input biosignals contain an index that is considered a dummy index. For example, CAD Severity-Overall, has 3 indexes that are correlated to the three possible mutually exclusive categories. The output can be given 4 indexes overall if a data point does not have a label for this output. During training if a biosignal does not have a label for the output, the dummy index is occupied, but with this loss function, the dummy index is ignored. The categorical cross-entropy is calculated and averaged along legitimate outputs. According to some non-limiting aspects, if the output does not have a corresponding label, the computer system 104 ( FIG. 1 ) can create a label on the fly that corresponds to the output that did not previously have a corresponding label. def f (y_true, y_pred): sum_cols=K.clip (K.sum (y true,axis=0), 0.0, 1.0) num_in=K.sum (sum cols [0: mask_value])+K. epsilon ( ) c_ce=K. categorical cross entropy (y_true, y_pred) loss=0.0 for i in range (mask value): inter mask=K. cast (K. equal (K. argmax (y_true), i), K. float x ( )) loss+=1.0/(num in)*K.sum (inter_mask*c_ce) ¥ /(K.sum (inter_mask)+K. epsilon ( )) return loss The loss function below can be used by the any output that is of the multi-label variety. Input biosignals that contain −1 can be masked and ignored. Binary cross-entropy can be calculated for both the true outputs and the predicted outputs. The cross-entropy values can be average for each of the indexes in the output and then the cross-entropy values for the 1 and 0 values. Only the indexes that exist in the true output are considered in the loss. def binary_loss (y_true, y_pred): mask_true=K. cast (K. not equal (y_true, mask_value), K. float x ( )) y_pred=mask_true*y_pred y_true=mask_true*y_true loc_0=mask_true*K. cast (K. equal (y_true, 0), K. float x ( )) loc_0_ve_c=K.sum (loc_0, axis=0) loc_1_ve_c=K.sum (y_true, axis=0) b_ce=K. binary_crossentropy (y_true, y_pred) loss_0=K.sum (loc_0*b_ce, axis=0)/(loc_0_ve_c+K. epsilon ( )) loss_1=K.sum (y_true*b_ce, axis=0)/(loc_1_ve_c+K. epsilon ( )) loss_0=K. mean (tf.boolean_mask (loss_0, tf. greater (loc_0_ve_c, 0))) loss_1=K. mean (t f.boolean_mask (loss_1, tf.greater (loc_1_ve_c, 0))) loss_0=tf. cond (tf. math. is_nan (loss_0), lambda: 0.0, lambda: loss_0) loss_1=tf.cond (tf. math.is_nan (loss_1), lambda: 0.0, lambda: loss_1) bottom=K. clip (K.sum (loc_O_ve_c),0.0,1.0)+¥ K. clip (K.sum (loc_1_ve_c), 0.0, 1.0)+K. epsilon ( ) loss=(loss_1+loss_0)/bottom return loss The loss function below is used by HEARTio for explicit semi-supervised learning. The concatenated feature vector is used to create a reference similarity matrix. Each output of HEARTio is then used to create a comparative similarity matrix. Signals that are similar across the feature vector space should be similar across the output space as well. Binary cross entropy is done between the reference similarity matrix and each output comparative matrix and then averaged. def get_mat (feature): r=tf.reduce_sum (feature*feature, 1) r=tf.reshape (r, [−1, 1]) D=−1*(r−2*tf.matmul (feature, feature, transpose_b=True)\+tf.transpose (r)) return (D−K.min (D)+1 e−7)/(K.max (D)−K.min (D)+1 e−7) def calculate_SSL_loss (reLmat, outpuLlist): num_outputs=len (output_list) A=get_mat (ref_mat) loss=0.0 for output in output_list: inter_mat=get_mat (output) loss+=1.0/num_outputs*\ K. mean (K. binary_crossentropy (A, inter_mat)) return loss def SSL_wrapper (ref_mat, output_list): return K.in_train_phase (calculate_SSL_loss (ref_mat, output_list),0.0) FIG. 9 is a block diagram 2400 of a HEARTio computer device, such as the computer device 104 ( FIG. 1 ), according to various aspects of the present invention. The illustrated computer device 2400 includes multiple processor units 2402 A-B that each includes, in the illustrated aspect, multiple (N) sets of processor cores 2404 A-N. Each processor unit 2402 A-B may include onboard memory (ROM or RAM) (not shown) and off-board memory 2406 A-B. The onboard memory may include primary, volatile, and/or non-volatile storage (e.g., storage directly accessible by the processor cores 2404 A-N). The off-board memory 2406 A-B may include secondary, non-volatile storage (e.g., storage that is not directly accessible by the processor cores 2404 A-N), such as ROM, HDDs, SSD, flash, etc. The processor cores 2404 A-N may be CPU cores, GPU cores and/or Al accelerator cores. GPU cores operate in parallel (e.g., a general-purpose GPU (GPGPU) pipeline) and, hence, can typically process data more efficiently that a collection of CPU cores, but all the cores of a GPU execute the same code at one time. Al accelerators are a class of microprocessor designed to accelerate artificial neural networks. They typically are employed as a coprocessor in a device with a host processor 2410 as well. An Al accelerator typically has tens of thousands of matrix multiplier units that operate at lower precision than a CPU core, such as 8-bit precision in an Al accelerator versus 64-bit precision in a CPU core. In various aspects, the different processor cores 2404 may train and/or implement different components of the HEARTio neural model. For example, in one aspect, the cores of the first processor unit 2402 A may implement the autoencoder, and the second processor unit 2402 B may implement the dense network, ResNet and/or Vit, as the case may be. In other aspects, one or more of the processor cores 2404 and/or one or more of the processor units could implement other components in the HEARTio neural model. One or more host processors 2410 may coordinate and control the processor units 2402 A-B. In other aspects, the system 2400 could be implemented with one processor unit 2402 . In aspects where there are multiple processor units, the processor units could be co-located or distributed. For example, the processor units 2402 may be interconnected by data networks, such as a LAN, WAN, the Internet, etc., using suitable wired and/or wireless data communication links. Data may be shared between the various processing units 2402 using suitable data links, such as data buses (preferably high-speed data buses) or network links (e.g., Ethernet). The software for the HEARTio model network and other computer functions described herein may be implemented in computer software using any suitable computer programming language, such as .NET, C, C++, or Python, and using conventional, functional, or object-oriented techniques. For example, the HEARTio neural mode may be implemented with software modules stored or otherwise maintained in computer readable media, e.g., RAM, ROM, secondary storage, etc. One or more processing cores (e.g., CPU or GPU cores) of the machine learning system may then execute the software modules to implement the function of the respective machine learning system (e.g., student, coach, etc.). Programming languages for computer software and other computer-implemented instructions may be translated into machine language by a compiler or an assembler before execution and/or may be translated directly at run time by an interpreter. Examples of assembly languages include ARM, MIPS, and x86; examples of high-level languages include Ada, BASIC, C, C++, C#, COBOL, Fortran, Java, Lisp, Pascal, Object Pascal, Haskell, ML; and examples of scripting languages include Bourne script, JavaScript, Python, Ruby, Lua, PHP, and Perl. The examples presented herein are intended to illustrate potential and specific implementations of the present invention. It can be appreciated that the examples are intended primarily for purposes of illustration of the invention for those skilled in the art. No particular aspect or aspects of the examples are necessarily intended to limit the scope of the present invention. Further, it is to be understood that the figures and descriptions of the present invention have been simplified to illustrate elements that are relevant for a clear understanding of the present invention, while eliminating, for purposes of clarity, other elements. While various aspects have been described herein, it should be apparent that various modifications, alterations, and adaptations to those aspects may occur to persons skilled in the art with attainment of at least some of the advantages. The disclosed aspects are therefore intended to include all such modifications, alterations, and adaptations without departing from the scope of the aspects as set forth herein. As previously discussed, deep learning, and machine learning models are employed by the HEARTio computer device disclosed herein to classify input biosignals such as electrocardiograms (ECGs) and electroencephalograms (EEGs), for example. Such biosignals can use signal-based methods such as recurrent neural networks, or long-term short-term memory models. HEARTio converts the biosignals into a multi-dimensional input matrix or “biosignal image” of the biosignal. HEARTio then processes the multi-dimensional input matrix/image to classify the initial biosignals. HEARTio is a multi-output model that has a single input, none of which are demographic or biometric based. HEARTio employs an autoencoder step that can “correct” or “fill-in” any incoming biosignal. This means if there is significant noise or artifact intrinsic in the input biosignal, or if any leads or parts of leads are missing, HEARTio can reconstruct a viable image, such as a 12 lead ECG system, for example. HEARTio can accomplish this with incomplete signal inputs. For example, HEARTio can reconstruct a 12 lead ECG even if there is only 1 viable input lead. Raw Outputs given by the HEARTio deep learning model include and LAD FFR, “70%” Disease threshold, Hypertrophy, Morphology, MI location, ST-T changes, and Conduction disorders. Training methodologies employed by the HEARTio deep learning model can include different ways of augmenting existing datasets. For example, stripping random segments/leads from the biosignal before inputting the biosignal during the forward pass segment of backpropagation. HEARTio can alter the biosignal, such as by increasing/decreasing the heart rate before input. Alternately and/or additionally, HEARTio can alter a sensor configuration, such as by flipping the leads before input, or reorganizing the leads before input. Adaptive weight losses changed after every epoch based on the change in performance and loss magnitudes. Training involves organizing each biosignal into a specific “folder” that is associated with a specific output. During training, HEARTio will be inputted with various “folders” such that at the end of a “cycle”, each folder has been shown to HEARTio so there is no gap in training. (i.e. a rare output that is only seen 1× every 500 epochs or something). During training, any ECGs that represent a “missing signal” with no useable information are dropped based on a frequency domain analysis. During training, any leads that represent a “missing signal” with no useable information are zero-ed out based on a frequency domain analysis Losses employed by the HEARTio deep learning model during training and evaluation. HEARTio has 4 major classes of custom-made loss functions that are continually weighed for each update. Loss functions include, for example, a weighed binary cross entropy, a weighed categorical cross entropy, a weighed log cosh, and a binned mean log squared error. Alternatively and/or additionally, HEARTio uses multiple semi-supervised loss functions in order to learn from data where there might not be a clinical output, just the ECG signal itself. For example, HEARTio can use a multi-task learning loss (MTL) that mandates that there is a correlation (either positive or negative) between the latent space and every one of the outputs. In other aspects, HEARTio can use a semi-supervised information loss (SSIL) that tries to have the model find an equilibrium between the information provided by each of the outputs. According to some non-limiting aspects, the semi-supervised loss and MTL losses can be added into an overall loss with some weighting. While the other losses (e.g., adaptive weight losses) can be used to train the model to make correct predictions on the training data, the semi-supervised, amd MTL can make sure that the model is generalized and not overfitting to the training data. As illustrated in FIGS. 5 A-B and 6 A-B, HEARTio can provide various visualizations employed by the platform. HEARTio can be device agnostic and configured to utilize any digital biosignal file coming from a biosignal device. If the only available file type is an image or pdf, a rudimentary computer vision algorithm can be used to convert this data type into a digital format. As such, HEARTio can use a full, 12-lead, ECG device or and single lead devices, including a wearable device, such as a smart watch. According to some non-limiting aspects, the present disclosure contemplates a diagnostic tool, wherein the diagnostic tool includes: a sensor for capturing at least one biosignal produced by a patient's heart; and a computer device that implements a deep neural network that is trained iteratively through machine learning to generate a prediction about a heart condition of the patient, wherein, after the deep neural network is trained, the computer device is configured to: convert the at least one biosignal to a multi-dimensional input matrix for the deep neural network generated from a number (N) of biosignals captured by the sensor, wherein each of the N biosignals is at least “T” seconds, and wherein N is greater than 1; and process the multi-dimensional input matrix through the deep neural network, wherein an output of the deep neural network from processing the multi-dimensional input matrix through the deep neural network corresponds to the prediction about the heart condition of the patient. In some non-limiting aspects, the deep neural network is a DenseNet including a plurality of layers configured to receive inputs from preceding layers of the plurality and generate outputs for proceeding layers of the plurality, such that each layer of the plurality has a collective knowledge from all preceding layers of the plurality. In some non-limiting aspects, the biosignal is an ECG signal, and wherein N is 12. In some non-limiting aspects, the biosignal is an ECG signal, wherein N is less than 12, and wherein the computer device further implements an autoencoder trained via machine learning to generate 12-N biosignals based, at least in part, on a reconstruction loss function. In some non-limiting aspects, the reconstruction loss function calculates a squared difference between a fast Fourier transform (FFT) between an input of the sensor and an output of the autoencoder. In some non-limiting aspects, the autoencoder includes: an encoder configured to perform a lossy compression of the at least one biosignal captured by the sensor, wherein an output of the lossy compression is a latent space representation; and a decoder configured to receive the latent space representation from the encoder and convert the latent space representation into a tensorform signal input to the DenseNet. In some non-limiting aspects, the DenseNet is configured to generate a feature vector based, at least in part, on the tensorform signal. In some non-limiting aspects, the DenseNet includes: a concatenation layer configured to concatenate the feature vector and the latent space representation into a concatenated feature vector; and a classifier clayer configured to generate outputs based, at least in part, on the concatenated feature vector, and wherein the prediction is generated based, at least in part, on the outputs of the classifier. In some non-limiting aspects, the output includes at least one of a beat classifications, a rhythm classification, and a coronary condition, or combinations thereof. In some non-limiting aspects, the coronary condition includes at least one of: a Myocardial Infarction (MI), a risk of a Major Adverse Cardiovascular Event (MACE), a coronary artery disease (CAD), a left anterior descending (LAD) coronary artery fractional flow reserve (FFR), an atherosclerotic cardiovascular disease (ACVD), cardiac hypertrophy, ventricle morphology, an abnormal ST-T wave, a conduction disorder, and a “70%” disease threshold, or combinations thereof. In some non-limiting aspects, wherein at least one biosignal is an electroencephalogram (EEG). In some non-limiting aspects, the sensor is a wearable device. In some non-limiting aspects, T is greater than or equal to 5 second and less than or equal to 15 seconds. In some non-limiting aspects, the diagnostic tool further includes a display coupled to the computer device, wherein the display is configured to visually represent the output of the deep neural network and the prediction about the heart condition of the patient. According to some non-limiting aspects, the present disclosure contemplates a method including: training, with a computer system, iteratively through machine learning, a deep neural network to make a prediction about a heart condition of a patient; and after training the deep neural network: capturing, with a sensor, at least one biosignal produced by a patient's heart; converting, by the computer system, the at least one biosignal to an input matrix for the deep neural network, wherein the input matrix includes a multi-dimensional matrix generated from a number (N) of biosignals of at least “T” seconds, where N is greater than 1; and processing, by the computer system, the multi-dimensional input matrix through the deep neural network, wherein an output of the deep neural network from processing the multi-dimensional input matrix through the deep neural network corresponds to the prediction about the heart condition of the patient. In some non-limiting aspects, the method further includes randomly removing at least one of the N biosignals from the multi-dimensional input matrix prior to processing the multi-dimensional input matrix through the deep neural network. In some non-limiting aspects, the method further includes altering either the biosignal captured by the sensor or a sensor configuration prior to processing the multi-dimensional input matrix through the deep neural network. In some non-limiting aspects, the processing is iterative, wherein the deep neural network includes a loss algorithm, and the method further includes: assessing, via the loss algorithm, a weight loss associated with each of the N biosignals in the multi-dimensional input matrix based, at least in part, on a change in performance and/or a loss magnitude; and changing the weight associated with each of the N biosignals for subsequent iterations based, at least in part, on the assessment. In some non-limiting aspects, the loss function is selected from a group consisting of a weighted binary cross entropy, a weighted categorical cross entropy, a weighted log cosh, and a binned mean log squared error. In some non-limiting aspects, the loss function is semi-supervised and is either a multi-task learning loss (MTL) configured to correlate latent space and a each of a plurality of outputs or a semi-supervised information loss (SSIL) configured to find an equilibrium between information provided by each the output and a plurality of possible outputs. In some non-limiting aspects, the method further includes organizing the sensor into one of a plurality of folders, wherein each of the plurality of folders is associated with at least one of the plurality of possible outputs. In some non-limiting aspects, the method further includes: assigning, by the computer system, a corresponding label to each of the plurality of outputs; determining, by the computer system, that at least one output of the plurality of outputs does not have a corresponding label; and creating, by the computer system, a label corresponding to the at least one output of the plurality of outputs. According to some non-limiting aspects, the present disclosure contemplates a computer system including one or more processor cores, wherein the one or more processor cores are configured to: train, iteratively through machine learning, a deep neural network to make a prediction about a heart condition of a patient; and after training the deep neural network: receive, from a sensor, at least one ECG signal produced by a patient's heart and captured by the sensor; convert the at least one ECG signal to an input matrix for the deep neural network, wherein the input matrix includes a multi-dimensional [or multi-axis] matrix generated from N ECG signals of at least T seconds, where N is greater than 1; and process the input matrix through the deep neural network, wherein an output of the deep neural network from processing the input matrix through the deep neural network corresponds to the prediction about the heart condition of the patient. The examples presented herein are intended to illustrate potential and specific implementations of the present invention. It can be appreciated that the examples are intended primarily for purposes of illustration of the invention for those skilled in the art. No particular aspect or aspects of the examples are necessarily intended to limit the scope of the present invention. Further, it is to be understood that the figures and descriptions of the present invention have been simplified to illustrate elements that are relevant for a clear understanding of the present invention, while eliminating, for purposes of clarity, other elements. While various aspects have been described herein, it should be apparent that various modifications, alterations, and adaptations to those aspects may occur to persons skilled in the art with attainment of at least some of the advantages. For example, functions explicitly described as performed by the previously disclosed diagnostic tool can be claimed as a method independent of the diagnostic tool, itself. Likewise, it shall be appreciated that the previously disclosed diagnostic tool can be further configured and claimed to perform steps explicitly described in reference to the methods disclosed herein. As such, the disclosed embodiments are therefore intended to include all such modifications, alterations, and adaptations without departing from the scope of the aspects as set forth herein.